Pulse Weekly #3, Anthropic's Busiest Week, Open Models Strike Back

2026-04-21 | 8 min read


What I found interesting from Hacker News, GitHub Trending, and Dev.to for the week of April 14–21, 2026. For HN, I looked at stories with 100+ points. For GitHub, I checked trending pages across 12 language categories (All, Go, Python, Rust, TypeScript, JavaScript, Java, C++, C, Shell, Kotlin, Swift) for daily, weekly, and monthly. For Dev.to, I pulled the top-reacted articles from the past 7 days. This is the first issue that includes Dev.to as a source.

Note: HN scores and GitHub star counts were captured on April 21, 2026. Dev.to reactions were captured on the same date. All numbers fluctuate over time.

#The Big Picture

Anthropic had the most packed week we’ve tracked: a new model, a design tool, a pricing change, an NSA controversy, and a $5B investment, all in 7 days. HN had opinions about every one. Three open-source coding models launched with direct comparisons to Claude. GitHub is still dominated by Claude Code ecosystem repos. Dev.to, the new source this week, shows a developer community more cautious about AI than GitHub trending suggests.

#1. Anthropic Had a Week

Anthropic shipped Claude Opus 4.7, launched Claude Design, removed Claude Code from the Pro plan, took $5B from Amazon, and dealt with reports of NSA using Mythos despite a blacklist, all in the same week. Each one landed on HN’s front page.

On HN:

PtsCommentsTitle
19551449Claude Opus 4.7
1230757Claude Design
712495Measuring Claude 4.7’s tokenizer costs
613574Anonymous request-token comparisons from Opus 4.6 and Opus 4.7
475337NSA is using Anthropic’s Mythos despite blacklist
472269Anthropic says OpenClaw-style Claude CLI usage is allowed again
383243Thoughts and feelings around Claude Design
314193Claude Code removed from Anthropic’s Pro plan
254261Anthropic takes $5B from Amazon and pledges $100B in cloud spending
11899Figma’s woes compound with Claude Design

On GitHub, Claude Code ecosystem repos dominate:

RepoStarsMonthlyWeeklyDaily
anthropics/claude-code, the CLI agent itself117K+36,565n/an/a
farion1231/cc-switch, multi-agent CLI for Claude Code, Codex, Gemini CLI48K+17,056+3,847+639
addyosmani/agent-skills, production-grade engineering skills for AI coding agents20Kn/a+3,452n/a
openai/openai-agents-python, lightweight multi-agent workflow frameworkn/an/a+3,078+600
coleam00/Archon, open-source harness builder for AI coding agents19K+5,389+1,572n/a

Opus 4.7 (1955 pts) got the most comments of any story this week (1449). Claude Design sparked two separate front-page discussions plus the Figma note, three threads about one product. Removing Claude Code from Pro generated two separate posts totaling 483 pts. The OpenClaw story is interesting context: last week Anthropic blocked OpenClaw-style usage, this week it allowed it again.

The Amazon deal is the quietest of the five: $5B investment with a pledge of $100B in AWS spending. 254 pts and 261 comments, but no GitHub spike tied to it.

#2. Open Source Models Took Aim at the Top

Three open-source coding models launched this week, all with explicit comparisons to Claude.

On HN:

PtsCommentsTitle
1270532Qwen3.6-35B-A3B: Agentic coding power, now open to all
998555Codex for almost everything
692357Kimi K2.6: Advancing open-source coding
690372Qwen3.6-Max-Preview: Smarter, Sharper, Still Evolving
46297Qwen3.6-35B-A3B on my laptop drew me a better pelican than Claude Opus 4.7
16546We got 207 tok/s with Qwen3.5-27B on an RTX 3090

On GitHub, agent frameworks beyond Claude Code are picking up:

RepoStarsMonthlyWeeklyDaily
NousResearch/hermes-agent, self-improving agent108K+95,651+30,630n/a
EvoMap/evolver, GEP-powered self-evolution engine for AI agents6.3K+4,358+4,032n/a
lsdefine/GenericAgent, skill-tree-based self-evolving agentn/a+4,183+3,914n/a
openai/openai-agents-python, multi-agent frameworkn/an/a+3,078+600
openai/codex, lightweight coding agent for the terminaln/a+10,424n/an/a

Qwen3.6-35B-A3B positions itself as a model that runs locally and handles agentic tasks. Kimi K2.6 from MoonshotAI targets coding performance directly. Codex from OpenAI is a terminal coding agent that got 998 pts and +10K monthly stars. Three different orgs dropped coding-focused models the same week Anthropic raised prices on Claude Code users.

hermes-agent added +30K stars this week, down from +48K last week, still massive but slowing. evolver and GenericAgent are both self-evolving agents with similar weekly growth. That’s a new pattern: two repos with the same architectural idea trending at the same time.

#3. Six Privacy and Security Stories Hit HN’s Top 10

Six privacy and security stories trended on HN this week, all unrelated to each other.

PtsCommentsTitle
1707763Google broke its promise to me – now ICE has my data
863487Vercel April 2026 security incident
791367GitHub’s fake star economy
766197Ban the sale of precise geolocation
590130Atlassian enables default data collection to train AI
398292€54k spike in 13h from unrestricted Firebase browser key accessing Gemini API

No GitHub repos trended this week around a privacy or security theme. This was an HN-only pattern.

The Google/ICE story (1707 pts) was the #3 story of the week. The Vercel breach (863 pts) generated four separate HN threads totaling over 1900 pts and 1500 comments, an OAuth attack through platform environment variables. The GitHub fake star economy piece (791 pts) covers how automated star inflation works. The Firebase Gemini key story (398 pts) covers a $54K charge from an unrestricted API key in client code.

#4. Agent Skills Are the New Plugin Store

The biggest pattern on GitHub this week: skills repos, configuration files and prompt collections for AI coding agents, are growing faster than the agents themselves.

RepoStarsMonthlyWeeklyDaily
forrestchang/andrej-karpathy-skills, single CLAUDE.md file72K+57,640+44,394n/a
farion1231/cc-switch, multi-agent CLI assistant48K+17,056+3,847+639
addyosmani/agent-skills, production-grade engineering skills20Kn/a+3,452n/a
SimoneAvogadro/android-reverse-engineering-skill, RE skills for Claude4.4K+2,591+2,299+483
Fission-AI/OpenSpec, spec-driven development for AI coding assistants42Kn/an/a+377

andrej-karpathy-skills went from 33K last week to 72K this week, +44K stars on a single text file. addyosmani/agent-skills is a skills collection from Addy Osmani, focused on production engineering tasks. cc-switch is in all three time periods for the second week in a row, no other AI repo has done that.

The format is maturing week over week. PW1 had resource hubs and awesome-lists. PW2 had memory plugins and workflow tools. PW3 has skills published under named engineers (Karpathy, Osmani). Personal reputation is now part of the distribution.

#5. Dev.to Debut: The Other Side of AI

This is the first issue to include Dev.to as a source. The top-reacted articles from the past 7 days cover a different angle than GitHub.

ReactionsTitle
121AI Doesn’t Fix Weak Engineering. It Just Speeds It Up.
67What Building with MCP Taught Me About Its Biggest Gap
66Don’t let AI do your thinking: a practical guide for engineers
49Lost in the AI Hype, I Started Small
39Stop Building AI Features Nobody Asked For

The same tone shows up on HN this week: The future of everything is lies, I guess (739 pts, 766 comments) and Laws of Software Engineering (794 pts) both made the top 10. The Dev.to top article says AI speeds up whatever engineering is already there, good or bad, a direct counterpoint to the skills wave on GitHub.

The Dev.to signal is different from HN. HN argues about the technology; Dev.to writes about living with it day to day. The MCP article (67 reactions) comes from someone building with MCP and running into a real gap, not debating the standard’s design.

#Quick Hits

#Week-Over-Week: PW2 (Apr 14) vs PW3 (Apr 21)

#My Take

The story of this week is the gap between what’s being built and what’s being discussed. GitHub has skills files from famous engineers and Claude ecosystem plugins. Dev.to and HN have engineers asking whether any of it makes their work better.

The pricing change (removing Claude Code from Pro) generated more HN comments than the model release. That’s a signal about where developer frustration is. andrej-karpathy-skills tripled its weekly growth, it’s a CLAUDE.md text file. The most starred category on GitHub this week is configuration, not code.

Three different orgs dropped open-source coding models the same week Anthropic pulled Claude Code from its cheaper plan. Data doesn’t tell you the intent, but the timing is there.


If this was useful, consider buying me a coffee :)